December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

Annu. Rev. Fluid. Mech. 1997. 29:473–514

Copyright c 1997 by Annual Reviews Inc. All rights reserved

UNSTRUCTURED GRID
TECHNIQUES
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

D. J. Mavriplis
Institute for Computer Applications in Science and Engineering NASA Langley
Research Center, Hampton, VA 23681-0001
by CAPES on 03/24/09. For personal use only.

KEY WORDS: computational, fluid, Navier-Stokes, aerodynamic, mesh, multigrid

ABSTRACT
An overview of the current state of the art in unstructured mesh techniques for
computational fluid dynamics is given. The topics of mesh generation and adapta-
tion, spatial discretization, and solution techniques for steady flows are covered.
Remaining difficulties in these areas are highlighted, and directions for future
work are outlined.

1. INTRODUCTION
The field of computational fluid dynamics has been evolving rapidly over the
past several decades. Aerospace applications, particularly in the field of aero-
dynamics, have provided the major impetus for the advances in computational
fluid dynamics technology. Great strides in available computational power,
coupled with rapid advances in algorithmic efficiency, have enabled computa-
tional fluid dynamics to substantially reduce the amount of wind-tunnel time
required for airframe and propulsion system design.
At the same time, computational methods are continuously being required
to deliver more accurate solutions for more complex realistic configurations at
lower computational cost. Traditionally, structured or block-structured meth-
ods that rely on regular arrays of quadrilateral or hexahedral cells in two or
three dimensions, respectively, have been employed to discretize the compu-
tational domain. Unstructured grid methods originally emerged as a viable
alternative to the structured or block-structured grid techniques for discretizing
complex geometries. These methods make use either of collections of simpli-
cial elements (triangles in two dimensions, tetrahedra in three dimensions) or
473
0066-4189/97/0115-0473$08.00
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

474 MAVRIPLIS

of elements of mixed type with irregular connectivity. This not only provides
greater flexibility for discretizing complex domains but also enables straightfor-
ward implementation of adaptive meshing techniques where mesh points may
be added, deleted, or moved about, while mesh connectivity is updated locally,
in order to enhance solution accuracy.
The advent of powerful desktop computers has greatly expanded the appli-
cability of computational fluid dynamics outside the realm of pure research
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

and aeronautics applications, into all areas of fluid mechanics, as well as into
multidisciplinary fields. The generality of unstructured grid methods and their
ability to enhance solution accuracy through adaptive procedures have proved
to be such a great advantage for these diverse applications that many if not most
commercial computational fluid dynamics codes currently rely on the use of
unstructured meshes.
by CAPES on 03/24/09. For personal use only.

Despite their recent success, unstructured mesh techniques still incur sub-
stantially more overhead than structured grid methods, a fact that has limited
their use particularly for large three-dimensional viscous flow cases. The devel-
opment of efficient and robust unstructured mesh algorithms for grid generation
and flow solution represents a considerable ongoing challenge. This paper pro-
vides an overview of the current state-of-the-art of such techniques; it points
out the advantages and deficiencies of various approaches as well as promising
avenues for future improvements.
In section 2 an overview of the most common grid generation and adaptivity
techniques is given. Section 3 describes the various approaches to spatial dis-
cretization for the Euler and Navier-Stokes equations, and section 4 describes
techniques for integrating the discretized equations to steady-state. Owing
to space limitations, the additional but equally important topics of unsteady
solution techniques and parallel implementations are not covered. Survey pa-
pers on these items can be found in the literature (see Special course 1992;
Venkatakrishnan 1995 for example).

2. MESH GENERATION
The initial phase of any numerical simulation begins with the generation of a
suitable mesh. The basic premise is that, owing to the random way elements
may be assembled to fill computational space, unstructured mesh generation
techniques are inherently more automatic and amenable to complex geometries
than traditional structured or block-structured grids. Therefore, the problem of
unstructured mesh generation is largely one of designing algorithms that are
automatic, robust, and yield suitable element shapes and distributions for the
flow solver.
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

UNSTRUCTURED GRID TECHNIQUES 475

In the following section, we examine the most prevalent techniques for gen-
erating unstructured meshes: the advancing-front method, and Delaunay-based
approaches. While the advancing-front method is somewhat heuristic in nature,
Delaunay-based methods are firmly rooted in computational geometry princi-
ples. In practice, both methods have resulted in successful three-dimensional
grid generation tools for arbitrary geometries. All of these techniques have also
demonstrated robustness problems and grid quality issues that have led to the
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

development of hybrid methods, which attempt to capitalize on the strengths

of the various methods.

2.1 Advancing-Front Techniques

In the advancing-front technique (Lo 1985, Peraire et al 1987, Gumbert et al
1989) an unstructured mesh is generated by adding individual elements one
by CAPES on 03/24/09. For personal use only.

at a time to an existing front of generated elements. Generation of a two-

dimensional grid begins with a discretization of the geometry boundaries as a
set of edges. These edges form the initial front that is to be advanced out into
the field. A particular edge of this front is selected, and a new triangle is formed
with this edge as its base by joining the two ends of the current edge either to
a newly created point or to an existing point on the front. The current edge
is then removed from the front, since it is now obscured by the new triangle.
Similarly, the remaining two edges of the new triangle are either assigned to
the front or removed from the front, depending on their visibility, as shown in
Figure 1.
The front thus constitutes a stack (or priority queue), and edges are con-
tinuously added to or removed from the stack. The process terminates when

Figure 1 Generation of a new triangle using new point (a), or existing front point (b), in the
advancing-front method. Dashed line edges denote current front.
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

476 MAVRIPLIS

the stack is empty, i.e. when all fronts have merged upon each other and the
domain is entirely covered. One of the critical features of such methods is the
placement of new points. Upon generating a new triangle, a new point is first
placed at a position that is determined to result in a triangle of optimal size
and shape. The parameters that define this optimum triangle as a function of
field position are obtained by a prescribed field function. (This field function
is often user defined and is evaluated by interpolating from values stored on a
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

coarse background grid.) The triangle generated with this new point may result
in a cross-over with other front edges, and may therefore be rejected. This
is determined by computing possible intersections with “nearby” front edges.
Alternately, an existing point on the front may coincidentally be located very
close to the new point, and thus should be employed as the forming point for
the new triangle, to avoid the appearance of a triangle with a very small edge at
by CAPES on 03/24/09. For personal use only.

some later stage. Existing candidate points are thus also searched by locating
all “nearby” front points.
For three-dimensional grid generation, a surface grid is first constructed by
generating a two-dimensional triangular mesh on the surface boundaries of the
domain. This triangular mesh forms the initial front, which is then advanced into
the flow-field by placing new points ahead of the front and forming tetrahedral
elements. The required intersection checking now involves triangular front
faces rather than edges as in the two-dimensional case.
Advancing-front methods generally result in smooth high-quality triangula-
tions in most regions of the flow-field. However, difficulties may be encoun-
tered in regions where the fronts must be merged. The algorithm is somewhat
heuristic in nature, and the need to perform intersection checking operations
on the moving front requires the use of dynamic quad/octree data structures for
efficient implementations.

2.2 Delaunay Triangulation Methods

Given a set of points in the plane, there exist many possible triangulations of
these points. A Delaunay construction represents a unique triangulation of these
points that exhibits a large class of well-defined properties (Preparata & Shamos
1985). The empty circumcircle property forms the basis of the Bowyer-Watson
algorithm (Bowyer 1981, Watson 1981). This property states that no triangle in
a Delaunay triangulation can contain a point other than its three forming vertices
within its circumcircle. Thus, given an initial triangulation, a new point may be
inserted into the triangulation by first locating and deleting all existing triangles
whose circumcircles contain the newly inserted point. A new triangulation is
then formed by joining the new point to all boundary vertices of the cavity
created by the previous removal of intersected triangles, as shown in Figure 2.
Point-insertion algorithms can be employed as the basis for a mesh generation
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

UNSTRUCTURED GRID TECHNIQUES 477

Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

Figure 2 Illustration of the Bowyer/Watson algorithm for constructing Delaunay triangulations.

by CAPES on 03/24/09. For personal use only.

strategy where the mesh points have been predetermined. The mesh points are
put in a list and an initial triangulation is artificially constructed (with auxiliary
points) that completely covers the domain to be gridded. The mesh points in
the list are then inserted sequentially into the existing triangulation using the
Bowyer-Watson algorithm. The final mesh is obtained when all points from the
list have been inserted. The mesh point distribution can in principle be random,
but most often it is either pregenerated using Cartesian mesh stencils (Baker
1987) or multiple overlapping structured meshes (Weatherill 1985; Mavriplis
1990a, 1991b) or may be generated on the fly using automatic point placement
strategies.
A simple automatic point placement technique has been proposed in Holmes
& Snyder (1988). Starting with an initial coarse triangulation that covers the
entire domain, a priority queue is constructed based on some parameter of the
individual triangles (circumradius, for example). A field value is assumed to
exist that determines the local maximum permissible value for the circumradius
of the triangles (or any other parameter). The first triangle in the queue is
examined, and a point is added at its circumcenter if the triangle circumradius
is larger than the locally prescribed maximum. This new point is inserted
into the triangulation using Bowyer-Watson’s algorithm, and the newly formed
triangles are inserted into the queue if their circumcircles are too large; otherwise
they are labeled as acceptable and do not appear in the queue. The final grid is
obtained when the priority queue empties out.
Delaunay triangulation techniques based on point insertion extend naturally
to three dimensions by considering the circumsphere associated with a tetrahe-
dron. These methods are more efficient than advancing-front techniques owing
to the simplicity of the algorithm. However, they are only valid on convex
domains. Therefore, Delaunay triangulation methods typically employ an arti-
ficially constructed convex domain (such as a quadrilateral in two dimensions or
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

478 MAVRIPLIS
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

Figure 3 Illustration of non-boundary-conforming triangulation for a simple airfoil geometry.

a cube in three dimensions) that covers the entire flow-field. The triangulation is
by CAPES on 03/24/09. For personal use only.

initially performed in this domain, and the triangles or tetrahedra that lie in the
flow-field are then extracted from the original convex domain to form the com-
putational mesh. The main disadvantages of Delaunay triangulation techniques
relate to their inability to guarantee boundary integrity. In two dimensions, the
set of edges that define the domain boundaries must form a subset of the edges
of the original convex domain triangulation. If this is not the case, edges of the
triangulation exist that penetrate the boundary, as shown in Figure 3.
In three dimensions, the equivalent problem is that of constructing a tetrahe-
dralization of the domain that contains, as a subset of its faces, a given surface
boundary triangulation. The use of local transformations as described in the
following section (edge and face-edge swapping, in two and three dimensions,
respectively), has been developed for locally modifying Delaunay meshes to
conform to specified boundary discretizations (George et al 1991, Weatherill
et al 1993).
In two dimensions, the existence of constrained Delaunay triangulations
(Chew 1989) guarantees the possibility of constructing triangulations that con-
tain a prescribed subset of edges and thus respect a given boundary discretiza-
tion. In three dimensions, the recovery of prescribed boundary triangulations
may not always be achievable through swapping techniques. A somewhat less
restrictive problem is that of constructing a tetrahedralization of the boundary
and interior points for which no edges or faces intersect the boundary surfaces.
One way of achieving this is by modifying the boundary point resolution (by
judiciously adding surface points) in regions where breakthroughs occur (Baker
1991).
2.3 Edge and Face Swapping Techniques
The Delaunay triangulation represents but one of many possible triangulations
of a given point set. An algorithm for transforming one triangulation to another
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

UNSTRUCTURED GRID TECHNIQUES 479

d
a
a
c
c b
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

2D e

3D

Figure 4 Two possible configurations for the diagonal of a convex pair of triangles in the edge-
by CAPES on 03/24/09. For personal use only.

swapping algorithm, and for two tetrahedra in the three-dimensional face-edge swapping algorithm.

triangulation of the same point set is given by Lawson (1972). Since all two-
dimensional planar graphs obey Euler’s formula (Mortenson 1985), all possible
triangulations of a given set of points contain the same number of edges and
triangles. Thus, any one triangulation may be obtained by simply rearranging
the edges of another triangulation of the same set of points. The edge-swapping
algorithm of Lawson (1972, 1977) consists of successively examining pairs of
neighboring triangles in the mesh. For each pair of triangles there exist two
possible configurations of the diagonal edge, as shown in Figure 4.
The algorithm chooses the diagonal configuration that optimizes some given
criterion. The algorithm terminates when no further optimizations are possible.
The edge-swapping algorithm can be used to transform an arbitrary triangula-
tion into a Delaunay triangulation. In this case, the edge-swapping criterion to
be used is the edge configuration that maximizes the minimum angle within the
two neighboring triangles. (Delaunay triangulations are also known as max-min
triangulations: They maximize the smallest angles in the triangulation.)
Edge-swapping can be used to construct triangulations other than the De-
launay triangulation, such as the minimum maximum angle (min-max) trian-
gulation, or the minimum total edge length triangulation (Preparata & Shamos
1985, Barth 1991a). However, the edge-swapping procedure is not guaranteed
to result in the global optimum triangulation for these cases: The algorithm
may terminate within local optima. A more sophisticated and complex algo-
rithm, (O(N 2 log N )), known as the edge-insertion algorithm (Edelsbrunner
et al 1992), has been developed for transforming a given triangulation into the
global optimum min-max triangulation.
In three dimensions, analogous local transformations can be achieved through
face- and edge-swapping. By considering pairs of neighboring tetrahedra, the
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

480 MAVRIPLIS

common triangular face may be removed and the edge joining the two oppos-
ing end points of the tetrahedra can be drawn, as shown in Figure 4. Note that
the original configuration that contained two tetrahedra is transformed into a
configuration containing three tetrahedra. Joe (1989) has shown that, when a
new point is introduced into an existing Delaunay tetrahedralization and con-
nected in an arbitrary manner, the global Delaunay tetrahedralization of the new
point set can be recovered through a sequence of face-edge swappings. How-
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

ever, the more general case of transforming an arbitrary tetrahedral mesh into
the Delaunay tetrahedral mesh through face-edge swapping often terminates
prematurely in a local optimum.
Face-edge swapping procedures are nevertheless useful in improving the
quality of three-dimensional meshes. In fact, Delaunay tetrahedralizations are
not necessarily the best constructions for three-dimensional meshes, since sliver
by CAPES on 03/24/09. For personal use only.

tetrahedra may easily be formed when four approximately co-circular points

are encountered. Therefore, face-edge swapping techniques based on the min-
max property are often employed to avoid or repair such ill-shaped elements,
even though the global min-max optimum for the entire grid is most often
unattainable.

2.4 Other Triangulation Methods

Quad/octree-based methods (Yerry & Shephard 1984) perform mesh generation
through a recursive subdivision of space down to a prescribed (spatially varying)
resolution. The vertices of the resulting quad or octree structure are used as
mesh points, and the tree quads or octants are divided up into triangular or
tetrahedral elements, in two or three dimensions, respectively. At boundaries,
the quad/octree cells intersect the boundary surfaces, and the vertices must be
displaced or warped to coincide with the boundary. The method is relatively
simple, inexpensive, and produces good quality meshes in interior regions of
the domain. The technique’s main drawback is an irregular cell distribution
near boundaries.
Various techniques have been developed that combine the advantages of both
advancing-front methods with the theoretical rigor of Delaunay triangulations.
Rebay (1993) and Mueller & Roe (1992) have developed similar algorithms,
where the triangulation method is based on Bowyer-Watson’s Delaunay trian-
gulation algorithm, but the point-insertion process is designed to mimic that
achieved by the advancing-front method. A background field function that de-
termines the desired distribution of element sizes is defined. A constrained De-
launay triangulation covering the entire domain and conforming to the bound-
aries is initially constructed. Mesh vertices are then inserted along the boundary
edges, at locations determined by the field function. This procedure results in
near-boundary elements that conform to the element size distributions specified
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

UNSTRUCTURED GRID TECHNIQUES 481

Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org
by CAPES on 03/24/09. For personal use only.

Figure 5 Partially completed advancing-front Delaunay point-insertion mesh.

by the background function. The boundary between these accepted elements

and other elements much larger than the required field function values consti-
tutes the front, and new mesh points are continuously added ahead of this front.
Figure 5 illustrates this process, where the front is clearly visible and is
advanced based on the minimum edge length in the front. The result is the
very smooth point distribution and high-quality mesh usually associated with
advancing-front techniques, coupled with the efficiency of the Delaunay point-
insertion algorithm, although the boundary integrity difficulties of the latter
method are inherited.
Marcum and Weatherill (1994) have developed a three-dimensional advanc-
ing-front point-insertion algorithm that makes use of face-edge swapping. An
initial boundary-conforming mesh is constructed, and new points are inserted
as described above. However, the Bowyer-Watson algorithm is not employed;
instead, the newly inserted vertex is simply connected to the four corners of
the existing element that contains it. A face-edge swapping operation is then
employed to improve the connectivity pattern of the elements in the vicinity of
the new vertex. One of the advantages of this approach is that the swapping
algorithm can be based on criteria other than the Delaunay triangulation. In
fact, the min-max criterion is most often employed in the reconnection strategy
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

482 MAVRIPLIS

because it avoids sliver tetrahedra. However, a straightforward application of

the min-max criterion quickly gets caught in local optima and yields poor-
quality meshes. A key feature that enables a closer approximation of the
global optimum and yields higher-quality meshes is the use of an intermediate
Delaunay-based swapping criterion. In this approach, each time a new point is
inserted into the mesh, the surrounding faces and edges are first swapped ac-
cording to the Delaunay triangulation criterion and then reswapped according
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

to the min-max criterion. The use of this intermediate construction serves to

broaden the range of influence of the new point, providing a more global ef-
fect and avoiding the immediate local optima that typically plague the straight
application of the min-max criterion.

2.5 Stretched Mesh Generation

by CAPES on 03/24/09. For personal use only.

The drive towards full Navier-Stokes solvers has necessitated the development
of stretched grid generation techniques in order to resolve the thin boundary-
layers, wakes, and other viscous regions characteristic of high-Reynolds-
number viscous flows. Proper boundary-layer resolution usually requires mesh
spacing several orders of magnitude smaller in the direction normal to the
boundaries than in the streamwise direction, resulting in large cell aspect-ratios
in these regions.
In Babushka & Aziz (1976) it is shown how the accuracy of a two-dimensional
finite-element approximation on triangular elements degrades as the maximum
angle of the element increases. Therefore, stretched obtuse triangles that con-
tain one large angle and two small angles are to be avoided, while stretched
right-angle triangles, with one small and two nearly right angles, are preferred.
Delaunay triangulations, which maximize the minimum angles of any trian-
gulation, tend to produce equiangular triangulations and are thus ill suited for
the construction of highly stretched triangular elements. One of the earliest ap-
proaches for generating highly stretched triangulations for viscous flows makes
use of a Delaunay triangulation performed in a locally mapped space (Mavriplis
1990b, 1991b; Vallet et al 1991; Castro-Diaz et al 1995). By defining a mapped
space based on the desired amount and direction of stretching, an isotropic De-
launay triangulation can be generated in this mapped space that, when mapped
back to physical space, provides the desired stretched triangulation. Difficulties
with such methods involve defining the stretching transformations and deter-
mining suitable point distributions for avoiding obtuse triangular elements.
The fact that obtuse triangles should be avoided has led to the development
of methods based on minimum-maximum angle triangulations in both two and
three dimensions (Barth 1991a, Marcum 1995). These can be implemented ei-
ther as a transformation from a Delaunay or other triangulation to a more optimal
min-max triangulation or by generating the min-max connectivity on the fly,
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

UNSTRUCTURED GRID TECHNIQUES 483

as points are inserted, using local reconnection techniques. Although the same
reconnection strategy can be employed in stretched and unstretched regions of
the flow-field, different point placement strategies must be employed in these
regions, in order to obtain the required anisotropic mesh-point resolution in
the stretched regions, and to avoid the formation of obtuse triangular elements,
which cannot be eliminated by reconnection strategies alone. Point-placement
strategies for the highly stretched regions most often mimic structured hyper-
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

bolic mesh generation techniques, where the points are placed along normal
lines emanating from each boundary vertex at intervals determined by the spec-
ified degree of mesh stretching. One of the drawbacks of min-max triangulation
approaches is that the resulting connectivity may be very sensitive to the dis-
tribution of vertices, with very small differences in mesh-point distributions
leading to substantially different connectivities.
by CAPES on 03/24/09. For personal use only.

An alternative to the above approaches is to generate a locally structured

or semistructured mesh in the regions where high stretching is required. One
approach (Nakahashi 1987, Ward & Kallinderis 1993) attempts to preserve the
mesh structure in the direction normal to the boundary up to a specified dis-
tance away from the boundary, after which fully unstructured isotropic meshing
techniques are employed. Special care must be taken in this case to avoid mesh
cross-overs in regions of concave curvature and to ensure a smooth transition
between the structured and unstructured region of the mesh. Another strategy
Löhner 1993; Pirzadeh 1994a,b; Connell & Braaten 1995) consists of gener-
ating a semistructured mesh, where the “stack” of mesh cells emanating from
each individual boundary face may terminate independently from those at other
boundary faces, as shown in Figure 6. Termination of these “advancing-layers”
(Pirzadeh 1994a) is triggered when the local cell aspect-ratio approaches unity,
or when cross-over with other cells is detected, such as in concave corners. The
remaining region is then gridded with a conventional isotropic unstructured
mesh generation approach. The resulting structured or semistructured meshes
can either be conserved as local structured entities of quadrilaterals in two di-
mensions and prisms in three dimensions (since the surface grid is generally
assumed to be triangular), or the different element types may be divided into
triangles or tetrahedra in two or three dimensions, respectively.

2.6 Mixed-Element Meshes

Delaunay constructions and face-edge swapping procedures are only relevant
for simplicial meshes (triangles in two dimensions, tetrahedra in three dimen-
sions). While simplicial elements provide the most flexibility in discretizing
complex geometries, the use of other types of elements may be beneficial. The
most common alternate element types are quadrilaterals in two dimensions, and
prisms, pyramids, and hexahedra in three dimensions.
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

484 MAVRIPLIS
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org
by CAPES on 03/24/09. For personal use only.

Figure 6 Illustration of the advancing-layers method for three-dimensional unstructured mesh

generation about segmented wing geometry. [Reproduced from Pirzadeh 1994b with permission.]
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

UNSTRUCTURED GRID TECHNIQUES 485

One of the advantages of nonsimplicial meshes is their lower connectivity.

For example, while tetrahedral meshes contain approximately six times as many
cells as vertices, and seven times as many edges as vertices, hexahedral meshes
contain the same number of cells as vertices, and only three times the number
of edges, while prismatic meshes contain twice as many cells and four times as
many edges (asymptotically, neglecting boundary effects). Since the cost of a
vertex-based discretization is directly proportional to the number of edges in the
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

mesh, a substantial efficiency gain can be realized by operating on nonsimplicial

meshes for a given number of unknowns or grid points. Furthermore, while
regular arrays of hexahedra or prisms can be employed to discretize three-
dimensional space, it is impossible to pack space with arrays of self-similar
tetrahedra. Therefore, regular arrays of nonsimplicial elements may also en-
hance accuracy, owing to a local cancellation of truncation errors that may not
by CAPES on 03/24/09. For personal use only.

occur on groups of nonsimilar tetrahedra. On the other hand, much of the

accuracy argument is either qualitative or speculative, and a controlled quan-
titative study has yet to be performed. For highly stretched two-dimensional
grids, Aftosmis et al (1994) observed similar accuracy on equivalent grids of
quadrilaterals and triangles in boundary layer regions and advocate the use of
quadrilateral elements owing to their lower connectivity.
Hybrid or mixed-element meshes may be constructed in a variety of manners.
For example, advancing-front techniques may be modified to generate quadri-
lateral or hexahedral elements (Blacker & Stephenson 1991, Zhu et al 1991),
although simplicial elements are often required to mend the mesh in regions of
geometric complexity. In regions of high mesh stretching, hexahedral or pris-
matic meshes are in some sense natural (Nakahasi 1987, Ward & Kallinderis
1993, Connell & Braaten 1995), depending on whether the surface discretization
contains triangular or quadrilateral elements, because the definition of a stretch-
ing direction implies a certain degree of structure in the mesh, which can be ex-
ploited using locally structured or semistructured mesh generation techniques.
Another approach is to attempt to reduce the complexity of an existing mesh
by combining simplicial elements together to form more complex elements
with lower connectivity (Merriam 1991, Tembulkar & Hanks 1992, Mavriplis
& Venkatakrishnan 1995b). Mesh merging techniques of this sort are most often
based on detecting groups of triangular or tetrahedral elements whose forming
vertices are nearly co-circular and which constitute a well-shaped quadrilateral
or prismatic element when merged. While merging algorithms have been rela-
tively successful in two dimensions, robust three-dimensional merging criteria
have yet to be demonstrated. A major advantage of the mesh-merging approach
is that the hybrid representation of the mesh remains internal to the flow solver,
which greatly simplifies other operations such as mesh adaptation and visual-
ization, which may operate on a fully simplicial representation of the mesh.
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

486 MAVRIPLIS

2.7 Adaptive Meshing

Aside from the treatment of complex geometries, the second main advantage
of unstructured meshes is the ease with which solution-adaptive meshing may
be implemented. Since no inherent structure is assumed in the representation
of the mesh, mesh points may be added, deleted, or displaced, and the mesh
connectivity may be locally reconfigured in the affected regions. The goal of
mesh adaptation is the determination of the optimum mesh-point distribution
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

that results in equipartition of the error for each individual simulation.

The character of the problem to be solved dictates the requirements of the
mesh adaptation strategy. For example, steady-state problems usually involve a
small number of adaptation phases as part of a lengthy solution process. There-
fore, relatively sophisticated adaptation strategies can be employed, such as, in
the extreme case, complete mesh regeneration. Mesh refinement procedures
by CAPES on 03/24/09. For personal use only.

are most important here, while de-refinement has only a minor effect and can
often be omitted for steady-state cases. For transient problems, mesh adapta-
tion must be performed every several time steps, and thus efficiency is much
more important than optimality. Mesh refinement and de-refinement are both
essential for transient cases, as well as mesh movement for cases with moving
boundaries. Furthermore, the accuracy of interpolation from the original mesh
to the refined mesh affects the solution accuracy (unlike the steady-state case),
and thus accurate transfer schemes are required.
Delaunay-based mesh generation techniques can easily be extended to in-
corporate adaptive refinement capabilities (Mavriplis 1990a, 1995b; Weatherill
et al 1993; Barth 1992). Once a solution has been obtained on an initial mesh,
new points can be added in regions where high errors are detected. These
new points can be triangulated into the mesh using the Bowyer-Watson point-
insertion algorithm. Alternatively, if a non-Delaunay mesh is employed, new
points may be inserted through element subdivision, and the connectivity of the
resulting mesh may be optimized through several face-edge swapping passes
based on any appropriate criterion.
Rule-based hierarchical element subdivision is a very effective adaptive tech-
nique, particularly for unsteady flows, where efficiency and accuracy of inter-
polation between successive grids are important considerations (Stoufflet et al
1987, Löhner & Baum 1992, Rausch et al 1992, Braaten & Connell 1996). The
essential approach consists in recursively subdividing mesh elements where
large solution errors are detected, but conforming to a set of well-defined sub-
division rules, which are necessary to prevent the formation of degenerate ele-
ment shapes and connectivities. Storing the hierarchy of the recursive process
enables de-refinement to be implemented by simply retracing the subdivision
history. Hierarchical subdivision is also applicable to nonsimplicial hybrid
meshes by constructing a library of subdivision rules for each element type.
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

UNSTRUCTURED GRID TECHNIQUES 487

Mesh movement or r-refinement has applications in both steady and un-

steady flows. For example, r-refinement may be useful for optimizing the
normal resolution of the mesh in a developing boundary layer or for clus-
tering points around shock waves (Palmerio 1994, Castro-Diaz et al 1995).
For transient problems with moving boundaries, r-refinement is indispensable.
Difficulties center around conserving a minimum degree of grid quality under
severe deformations. In areas where grid deformation becomes unacceptable,
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

local reconnection through swapping techniques can easily be performed, but

only for simplicial meshes (Kennon et al, Baum et al 1994, Venkatakrishnan &
Mavriplis 1995b). In practice, combinations of h- and r-refinement are often
employed (Baum et al 1994, Castro-Diaz et al 1995) for such cases.
Central to the implementation of any solution-adaptive scheme is the ability
to detect and assess solution error. The construction of a suitable refinement
by CAPES on 03/24/09. For personal use only.

criterion represents the weakest point of most adaptive strategies. The main
problem is that an exact characterization of the error requires a knowledge of
the solution itself, which is obviously impractical. Most error estimates are
based on the assumption that the solution is smooth and asymptotically close
to the exact solution. This is often not the case for fluid dynamics problems,
which are governed by nonlinear hyperbolic partial-differential equations. So-
lutions may contain discontinuities, and downstream flow features often depend
on adequate resolution of upstream flow features. Most refinement criteria are
heuristically based on (undivided) gradients and/or second differences of the
various flow variables. Conservative criteria (i.e. over-refining) are often em-
ployed to compensate for the inability to accurately characterize the true solution
error. Because adaptive meshing results in different mesh topologies for each
simulation, even when the geometry of the problem is unchanged, parametric
studies (typically used in design processes) are complicated by the requirement
to distinguish between grid-induced and physical solution variations. Never-
theless, for problems with disparate length scales, adaptive meshing is often
indispensable for resolving small flow features, and their full potential awaits
the development of more well-founded adaptive criteria.

3. SPATIAL DISCRETIZATION
The previous section describes techniques for generating suitable meshes about
arbitrary geometric configurations. Once such a mesh has been generated, it
serves as the basis for the spatial discretization of the governing fluid dynamic
equations. Unstructured mesh discretization techniques can generally be clas-
sified as finite-volume or finite-element strategies, depending on whether a
discrete integral or variational viewpoint is adopted, although many common
discretizations may be simultaneously interpreted from both viewpoints.
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

488 MAVRIPLIS

LR
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

LR
by CAPES on 03/24/09. For personal use only.

Figure 7 Illustration of dual mesh for mixed triangular-quadrilateral unstructured mesh and as-
sociated control-volumes with edge-based fluxes for a vertex-based scheme.

For finite-volume formulations, a distinction between vertex-based and cell-

centered schemes can be made. In a vertex-based scheme, the flow variables
are stored at the vertices of the mesh. The discrete equations usually involve
stencils that are defined by the nearest or next-to-nearest neighbors of each
mesh point. In a cell-centered scheme, the flow variables are stored at the cen-
ters of the cells or elements of the mesh. The stencils of the discrete equations
usually involve the nearest or next-to-nearest neighboring cells. The equiv-
alence between vertex-based and cell-centered schemes can be demonstrated
with the concept of a dual mesh. If a dual mesh point is created at each cell
center and dual mesh edges are drawn by joining neighboring cell centers, the
cell-centered scheme can be seen to be equivalent to a vertex-based scheme
operating on the dual mesh.
Figure 7 illustrates the dual mesh for a mixed quadrilateral and triangular
mesh in two dimensions. For a purely quadrilateral or hexahedral mesh, the
dual mesh also contains quadrilateral or hexahedral elements, and the number
of vertices and edges in the original and dual meshes is identical. For triangular
or tetrahedral meshes, the number of dual mesh vertices is larger than the
number of original mesh vertices. This is due to the fact that a triangular mesh
contains twice as many triangles as vertices (neglecting boundary effects), and
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

UNSTRUCTURED GRID TECHNIQUES 489

a tetrahedral mesh five to six times more elements than vertices. On the dual
mesh, the degree of a vertex (number of incoming edges) is fixed and equal to
three for triangular elements, or four for tetrahedral elements, whereas on the
original mesh, the degree of each vertex is variable. Similar relationships hold
for other elements such as prisms and pyramids.
Thus, cell-centered and vertex-based schemes operating on the same grid
result in vastly different discretizations when nonquadrilateral or nonhexahedral
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

elements are present. In particular, a cell-centered scheme can be expected to

result in a much larger number of unknowns while generating relatively simple
stencils of fixed size. The vertex-based scheme, on the other hand, will result
in a smaller number of unknowns with larger variable-size stencils. Because of
the larger number of unknowns, cell-centered schemes generally incur larger
overheads than vertex-based schemes on equivalent grids. However, there is
by CAPES on 03/24/09. For personal use only.

evidence to suggest that they also provide more accurate solutions on equivalent
grids. The question of whether the additional overheads are offset by the
increase in accuracy is still an open one, especially since vertex-based schemes
operate on more complex stencils (more fluxes per unknown) than cell-centered
schemes.
3.1 Central Differencing and Riemann-Based
Upwind Methods
A simple finite-volume discretization can be obtained by defining a control
volume about each unknown (mesh-element for cell-centered schemes, dual-
mesh cell for vertex-based schemes) and integrating the fluxes over the boundary
of the control volume using a simple midpoint integration rule, where the value
of the flux at each control volume face, as shown by the arrows in Figure 7, is
computed by averaging the flow variables in the two control volumes on either
side of the face. For vertex-based schemes on simplicial meshes, this particular
scheme can also be derived using Galerkin finite elements, where the fluxes are
assumed to vary linearly over the triangular or tetrahedral elements of the mesh
(Jameson et al 1986, Rostand & Stoufflet 1988, Barth 1991a). Such a scheme
corresponds to a central difference discretization on structured meshes and
requires the addition of artificial diffusive terms to ensure stability. By analogy
with the blended second and fourth differences employed on structured meshes
(Jameson et al 1981), artificial dissipation is constructed as a combination of
an undivided Laplacian and biharmonic operator. First-order shock-capturing
diffusive terms are provided by the Laplacian operator, which is turned off away
from shocks in regions of smooth flow where the biharmonic dissipation ensures
stability with second-order accuracy. This approach has been used extensively
to compute inviscid and viscous two- and three-dimensional flows (Jameson
et al 1986; Mavriplis 1987, 1992; Smith 1990; Mavriplis & Venkatakrishnan
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

490 MAVRIPLIS

1995a; Peraire et al 1992). The discretization is simple to implement and

relatively inexpensive to evaluate. However, the construction of the dissipative
terms is nonoptimal in that it does not differentiate between the various wave
components of the governing fluid dynamic equations.
Upwind schemes attempt to remedy this deficiency by applying the appro-
priate amount of dissipation to each wave component. This is achieved by di-
agonalizing the system of governing equations, using a local one-dimensional
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

approximation normal to each control-volume interface, and constructing the

dissipation terms in the diagonalized or “split” form of the equations. In prac-
tice, this is done by solving a Riemann problem at the interfaces between
neighboring control volumes. For first-order schemes, the flow variables are
assumed to be piecewise constant over the control volumes. The values on ei-
ther side of the interface, often referred to as “left” and “right” values (denoted
by CAPES on 03/24/09. For personal use only.

by L and R in Figures 7 and 8), are then simply taken as the values in the neigh-
boring control volumes. For second-order approximations, the flow variables
are assumed to vary linearly over the control volumes, and interface values
must be extrapolated from centroidal values to the control-volume interfaces
using estimated gradient information. Higher-order methods may similarly be

L R
j
i

Figure 8 Illustration of neighboring triangles employed to compute gradients in direction of edge

i j for schemes in Desideri & Dervieux 1988 and Jameson 1994. L and R refer to the left and
right states about the control-volume interface associated with edge i j for the scheme of Desideri
& Dervieux (1988).
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

UNSTRUCTURED GRID TECHNIQUES 491

formed by reconstructing approximations to higher-order derivatives in the so-

lution variables, which are then used to extrapolate more accurate values to
either side of the interface. The resulting interface values are discontinuous,
although in the absence of shock waves the magnitude of the jumps decreases
with increasing order of accuracy of the scheme. A Riemann solver is then
required to calculate a flux at the interface given the discontinuous left and
right states. Approximate Riemann solvers are typically employed, the most
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

common being those of Osher (1984) and Roe (1981). In order to prevent os-
cillations near discontinuities such as shocks, the extrapolated interface values
must be limited, such that the creation of new extrema is never permitted.
Some of the earliest applications of upwind schemes to unstructured meshes
were reported by Desideri & Dervieux (1988). They developed a vertex-based
scheme that estimates gradients along an edge from the projection of gra-
by CAPES on 03/24/09. For personal use only.

dients from neighboring triangles, as depicted in Figure 8, and employs an

Osher Riemann solver. Limiting was applied along cell interfaces in a quasi-
one-dimensional manner. Barth & Jespersen (1989) proposed a second-order
scheme, using a Roe Riemann solver, where both reconstruction and limiting
are done in a truly multidimensional fashion. Interface values are extrapolated
from the control-volume centroids by computing a gradient at each control
volume using a Green-Gauss integration around the control volume boundary.
These gradients are then limited to avoid the formation of new extrema, based on
neighboring values. The method was later extended to include a least-squares
technique for estimating gradients (Barth 1991a), which was shown to produce
better approximations on distorted meshes. Extensions using high-order re-
constructions have also been developed (Barth 1991b). A cell-centered upwind
scheme for three dimensional tetrahedral meshes has been developed by Frink
(1992) which achieves linear reconstruction by computing cell-based gradients
using vertex values obtained by a weighted average of surrounding cell-based
quantities. One of the drawbacks of upwind schemes is the required use of a
nonlinear limiter to capture discontinuities such as shock waves without numer-
ical oscillations. Many limiter formulations have been developed, but most are
sensitive to small variations in the solution variables, which in turn adversely
affect convergence of the overall scheme. Venkatakrishnan (1995b) has inves-
tigated this matter and has proposed a weaker form of limiter which permits
small oscillations which vanish as the grid is refined, but provides superior
convergence characteristics.
The development of upwind schemes for unstructured meshes has enabled the
re-evaluation of central difference schemes with added dissipation as a scalar
counterpart to upwind schemes. This has resulted in alternate formulations
based on gradients rather than Laplacian and biharmonic operators for the
artificial dissipation terms (Peraire et al 1992).
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

492 MAVRIPLIS

3.2 Other Approaches

A discretization scheme is said to be positive if, in the scalar case, it can be
expressed in the form
dwi X
= Ci j (w j − wi ) (1)
dt j
where the Ci j coefficients are all positive. The summation is over the neigh-
bors j of i. If wi is a maximum, then w j − wi is negative and dw < 0,
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

i
dt
from the positivity of the Ci j . Thus maxima must be decreasing, and con-
versely minima must be increasing. This Local Extremum Diminishing (LED)
principle (Jameson 1994) represents a multidimensional generalization of the
Total Variation Diminishing (TVD) concept (Harten 1984). This has important
implications for the construction of robust and accurate schemes. The LED
by CAPES on 03/24/09. For personal use only.

principle can be used to guarantee the capture of shock waves without any os-
cillations. Furthermore, the maximum principle inherent in these schemes can
provide a robust solution strategy—for example, by guaranteeing the positivity
of all flow quantities simply if they are initialized as positive quantities.
Unfortunately, if the Ci j’s are constant, the scheme is linear and can be at
best first-order accurate. To achieve second-order accuracy, a nonlinear scheme
must be constructed, where the Ci j coefficients depend on the solution variables.
These usually include the use of limiters based on ratios of successive gradients
on neighboring mesh intervals. These methods have their roots in the early
work of Boris & Book (1973) on flux-corrected transport schemes. The original
idea is to construct a first-order scheme, the accuracy of which is then refined
by adding specific amounts of antidiffusive fluxes, under the constraints of
monotonicity. This scheme was extended to unstructured meshes by Löhner in
the FEM-FCT schemes (Löhner et al 1987). Jameson has derived a Symmetric
Limited Positive (SLIP) scheme that is both positive and second-order accurate
in smooth regions for unstructured meshes (Jameson et al 1986), where one-
dimensional solution variations along stencil edges are obtained by projecting
adjacent triangle gradients into the direction of the edge, in a manner similar to
that of Desideri & Dervieux (1988) (cf Figure 8). These schemes can be applied
to the unsplit form of the governing flow equations, in a scalar dissipation
mode, or to the split equations, characteristically or otherwise, to achieve more
accurate resolution of shock waves and boundary layers. However, for systems
of equations, monotonicity is not strictly guaranteed, as it is in the scalar case.
Additionally, these schemes rely on the use of a nonlinear limiter in all regions
of the flow-field (not only in the vicinity of shock waves), which may adversely
impact convergence, depending on the form of limiter employed.
The weak link in the upwind schemes described in the previous section in-
volves the use of a dimensionally split Riemann solver, which is applied in
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

UNSTRUCTURED GRID TECHNIQUES 493

a quasi-one-dimensional fashion normal to control-volume interfaces. While

this approach is employed routinely for structured and unstructured grids, it
may misinterpret flow features not aligned with the control-volume interfaces.
Fluctuation splitting schemes provide an alternative to Riemann solvers for
introducing upwinding. In this approach, variables are stored at the mesh ver-
tices, and intermediate residuals are computed at cell centers. The values at
the vertices are then updated by redistributing the cell-based residuals to the
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

vertices. If these residuals are redistributed in an unbiased manner, then the

well-known Ni scheme (Ni 1982) is recovered, which corresponds to a central
difference scheme and requires additional dissipative terms. Upwind schemes
are obtained by introducing downwind biasing into the redistribution formulae
by considering the effect of linear wave solutions evolving the piecewise linear
data over a mesh cell. Several upwind fluctuation-splitting schemes for the
by CAPES on 03/24/09. For personal use only.

scalar advection equation have been developed and are compared in Paillere
et al (1993). Extensions of fluctuation-splitting schemes to systems of equa-
tions rely on the decomposition of the flux divergence into scalar waves. The
decomposition is not unique, and several strategies such as characteristic de-
composition (Deconinck et al 1986), simple wave models (Parpia & Michalak
1993, Roe 1986), and algebraic approaches (Sidilkover 1994) have been in-
vestigated. The advantages of fluctuation-splitting schemes are that they do
not require the use of dimensionally split Riemann solvers and they result in
a compact nearest-neighbor stencil for simplicial meshes. However, they are
second-order accurate only at steady state (Venkaktkrishnan 1996).
In the finite-element approach, the solution is expanded in a set of basis func-
tions, and the discrete equations are obtained by requiring the residuals to be
orthogonal to a set of trial functions. For second-order approximations, the basis
functions are generally taken as linear functions of compact support, which have
the value 1 at a given vertex and vanish at all other vertices. The solution expan-
sion is obtained by multiplying these basis functions by vertex-based-solution
variables and summing over the entire grid. When the test functions and the ba-
sis functions are the same, the method is called Galerkin, otherwise it is called
Petrov-Galerkin. The correspondence between (a) a Galerkin finite-element
strategy using linear elements and (b) finite-volume based central-differencing
has already been discussed. For the convective terms, additional artificial dis-
sipation is required for stability. The Galerkin finite-element method can also
be employed to discretize the viscous terms for the Navier-Stokes equations,
as discussed in the next section.
The Streamwise Upwind Petrov Galerkin (SUPG) (Hughes & Brooks 1979)
and Galerkin least-squares methods (Hughes et al 1989, Hughes 1987) intro-
duce dissipation to a regular Galerkin method by adding terms proportional
to the residual. In contrast to traditional artificial-dissipation formulations,
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

494 MAVRIPLIS

these approaches do not compromise the accuracy of the scheme, since at

steady-state convergence, the dissipation vanishes along with the residual. The
1
global error in the L 2 norm can be shown to vary as h p+ 2 for the advec-
p+1
tion dominated case and as h for the diffusion dominated case, where
p is the degree of the piecewise-polynomial basis function (Hughes 1987,
Venkatakrishnan 1996). However, these methods are linear and not suitable
for capturing discontinuities. A shock-capturing operator that locally reduces
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

the order of accuracy near discontinuities has been proposed by Hughes &
Mallet (1986).

3.3 Considerations for the Navier-Stokes Equations

The above discussion on discretizations is principally concerned with the purely
convective terms that arise in the context of inviscid flows. With the excep-
by CAPES on 03/24/09. For personal use only.

tion of the finite-element methods described above, where convection and

diffusion terms can be treated in a unified manner, additional viscous terms
must be discretized for the Navier-Stokes equations. These terms are dif-
fusive in nature and generally take the form of second differences. Stable
second-order discretizations can therefore be constructed using simple central
differences.
For cell-centered schemes, one approach consists of first computing the
gradients at the mesh vertices and then averaging these to the cell faces in
order to compute second derivatives at cell centers (Frink 1994). For vertex-
based schemes, a similar two-pass procedure based on finite-volume central-
difference arguments can be used to construct the viscous term discretization.
For simplicial meshes, gradients can first be computed on the mesh elements,
which can then be interpolated to the control-volume faces in order to form
second differences at the mesh vertices. This type of discretization can be
derived more formally using a Galerkin finite-element procedure and assumes
linear variation of the flow variables on the mesh elements. The resulting dis-
cretization produces a nearest-neighbor stencil and may be implemented as
a single loop over the edges of the mesh, rather than as a two-pass proce-
dure that computes intermediate cell-based gradients (Barth 1992, Mavriplis
1995c). For nonsimplicial meshes, Galerkin finite-element formulations us-
ing bilinear or trilinear variations on quadrilateral or hexahedral elements can
be constructed. The resulting stencils, however, are no longer compact, as
they involve vertices within mesh elements that are not connected by a mesh
edge, such as diagonally opposed vertices in hexahedral elements (Braaten &
Connell 1996).
An alternative strategy for discretizing viscous terms for vertex-based
schemes is to employ the vertex-based gradients already computed in the context
of second-order upwind schemes (using a Green-Gauss integration around the
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

UNSTRUCTURED GRID TECHNIQUES 495

vertex-based control volumes, for example), instead of the element-based gradi-

ents described above. A vertex-based second difference can then be computed
by integrating these gradients themselves around the control-volume bound-
aries (Luo et al 1993). This approach enables the viscous term discretizations
to be assembled on meshes of arbitrary element types using the same data struc-
tures as required for the upwind convection terms. The principal drawback of
this method is that it results in a large stencil that involves neighbors and next-
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

to-neighbors, which on a structured mesh reduces to a stencil of size 2h, where

h represents the mesh spacing. This not only reduces overall accuracy but is
also ineffective at damping odd-even oscillations in the numerical scheme.
For high-Reynolds-number flows of practical interest, the Reynolds-averaged
form of the Navier-Stokes equations is generally employed, which requires
the use of additional turbulence-modeling equation(s). Although algebraic
by CAPES on 03/24/09. For personal use only.

models can be implemented on unstructured grids (Mavriplis 1991a), they were

conceived for simple wall-bounded flows and are thus ill suited for flows over
complex geometries. The current practice is to employ two-equation models of
the k − or k −ω type, or simpler one-equation eddy viscosity models (Baldwin
& Barth 1992, Spalart & Allmaras 1992). These equations contain convective,
diffusive, and source terms that can be discretized in a manner analogous to
the discretization of the flow equations. Turbulence modeling equations often
result in stiff numerical systems, and care must be taken to devise schemes
that preserve positivity of the turbulence quantities at all stages of the solution
process. A common practice is to discretize the convective terms using a first-
order upwind strategy, from which a positive scheme that obeys a maximum
principle can be obtained.

3.4 Data Structures

A brief discussion on data structures is useful because the success of most
discretization methods ultimately depends on how efficiently they may be im-
plemented. Traditionally, finite-element methods have relied on element-based
data structures, where for each element of the mesh a list of the forming vertex
addresses is stored (i.e. four vertices for tetrahedra, eight vertices for hexahedra,
etc). For many fluid dynamics problems, the discretizations, which are typically
thought of as summations of fluxes, can be implemented more effectively using
an edge-based data structure (Barth 1992; Mavriplis 1992, 1995c; Peraire et al
1992; Luo et al 1993). For a vertex-based scheme, this corresponds to storing,
for each edge of the mesh, the addresses of the two vertices on either end of
the edge. For a cell-centered scheme, the relevant entity is the dual edge that
joins two neighboring cell centroids and pierces the face common to these two
cells. The discretization may be evaluated by computing a flux on each edge,
which is then added to and subtracted from the respective control volumes on
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

496 MAVRIPLIS

each end of the edge. In order to compute this flux, a face area must be stored
for each edge, which corresponds to the area of the dual control-volume face
associated with the mesh edge in the vertex-based scheme, and to the area of
the cell face pierced by the dual edge in the cell-centered scheme. The use of
edge-based data structures results in lower memory overheads and increased
computational throughputs because redundant computations are eliminated and
the gather-scatter required for vectorization in supercomputers is minimized.
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

Furthermore, because sets of edges can be used as building blocks for arbitrarily
shaped elements, hybrid meshes with mixed element types may be handled by
a single edge-based data structure, at least for inviscid flows.
For viscous flows, the Galerkin finite-element discretization of the diffusion
terms on simplicial meshes results in a nearest-neighbor stencil and thus may
be implemented using an edge-based data structure (Mavriplis 1995c). How-
by CAPES on 03/24/09. For personal use only.

ever, this is not the case for nonsimplicial meshes, since the resulting stencils
involve vertices that are not connected to the center vertex by a mesh edge
(such as diagonally opposed vertices in hexahedral elements). In these situ-
ations, the element-based data structure must be retained (Braaten & Connell
1996). An alternative is to resort to the thin-layer approximation of the vis-
cous terms on nonsimplicial meshes, which can be implemented exclusively
along edges (Mavriplis & Venkatakrishnan 1995b). The limitations of this
approach are obvious, although it is justifiable for highly stretched prismatic
or hexahedral meshes, where streamwise resolution has been sacrificed for
efficiency.
An interesting property of the edge-based data structure is that it can pro-
vide an interpretation of the discrete operator as a sparse matrix. For nearest-
neighbor stencil discretizations, all points in the stencil are joined to the center
point by a mesh edge. The discretization operator can be written as a sparse
matrix, where each nonzero entry in the matrix corresponds to a stencil coef-
ficient or edge of the mesh. For systems of equations, the edges correspond
to nonzero block matrix entries in the large sparse matrix. This interpretation
has implications for the implementation of implicit and algebraic multigrid
solution schemes (Venkatakrishnan & Mavriplis 1993, Mavriplis & Venkata-
krishnan 1995b).
One of the disadvantages of the edge-based data structure is that it requires
a preprocessing operation to extract a unique list of edges from the list of mesh
elements and to compute the associated edge coefficients. For unsteady flows
with dynamic meshes, this preprocessing must be performed every time the
mesh is altered, although this may be done locally. Additionally, for dynamic
grid cases, the element data structures are generally required for performing
mesh motion or adaptation, since edge lists represent a lower-level description
of the mesh.
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

UNSTRUCTURED GRID TECHNIQUES 497

4. SOLUTION TECHNIQUES
The discretization of the spatial terms in the governing equations results in a
large coupled set of ordinary differential equations of the form
d(V Mw)
+ R(w) = 0 (2)
dt
where V represents the local control volume, M the mass matrix, and R(w)
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

the spatially discretized terms. For a vertex-based scheme, the mass matrix
relates the average value of a control volume to the values at the vertices of
the mesh. Both V and M are constants for static meshes, and can therefore
be taken outside of the time derivative. For steady-state cases, these equations
must be integrated in time towards t → ∞, where the time derivatives become
vanishingly small. Since time accuracy is not a concern in such cases, the mass
by CAPES on 03/24/09. For personal use only.

matrix may be lumped—i.e. replaced by the identity matrix—and each equation

may be advanced with the maximum permissible time step, as determined from
local stability considerations.

4.1 Explicit Schemes

A simple time integration scheme is obtained by replacing the time derivative
by a simple forward difference and evaluating the residual at the current time
level:
wn+1 − wn
V t + R(wn ) = 0 (3)
1
This corresponds to a single-stage explicit scheme, since updates are obtained
from one evaluation of currently available quantities. Like Runge-Kutta
schemes for ordinary differential equations, multistage time-stepping schemes
(which are required with higher-order discretizations for stability reasons) in-
volve the combination of updates obtained at multiple explicitly evaluated in-
termediate states.
The stage coefficients of these schemes can be optimized either to provide
large time steps at the expense of time accuracy or to enhance high-frequency
damping properties of the scheme, which is required in the context of a multigrid
algorithm (Jameson et al 1981, van Leer et al 1989).
These schemes are extremely simple to implement and require little addi-
tional computer storage. However, stability considerations restrict the maxi-
mum permissible time step to values proportional to the local cell size. Thus,
finer meshes lead to smaller time steps, which in turn lead to larger solution
times. For very fine meshes, the convergence of explicit schemes becomes
unacceptably slow, and more sophisticated solution strategies must be adopted.
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

498 MAVRIPLIS

4.2 Implicit Schemes

An implicit scheme is obtained by evaluating the spatial residual terms at the new
time level n + 1. Since these quantities are not known explicitly, a linearization
must be performed about the current time level:
µ ¶
V ∂R
+ 1w = −R(wn ) (4)
1t ∂w
∂R
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

∂w
represents the Jacobian and constitutes a large sparse matrix. At each
time step, the above linear system must be solved for the corrections 1w =
(wn+1 − wn ) from which the flow variables can be updated.
Implicit schemes may be classified by the degree to which the true Jacobian
matrix is approximated. If an exact linearization is employed, the method is
unconditionally stable and reduces to Newton’s method for 1t → ∞. Al-
by CAPES on 03/24/09. For personal use only.

though the quadratic convergence properties of Newton’s method generally

produce solutions in a very small number of time steps (often of the order of
10) (Venkatakrishnan & Barth 1989, Barth & Linton 1995, Nielsen et al 1995),
each time step requires the inversion of a large sparse matrix, which becomes
prohibitively expensive in terms of storage and CPU-time for fine meshes.
A common simplification is to replace the exact linearization with one based
on a first-order discretization (Struijs et al 1989, Fezoui & Stoufflet 1989,
Venkatakrishnan & Mavriplis 1993, Anderson et al 1995). While this consid-
erably reduces the storage requirements of the linear system and improves its
condition number, it precludes attaining quadratic convergence rates, owing to
the use of an inexact linearization. This in turn favors the use of iterative meth-
ods to solve the linear system, which can be used to converge the system only
partially, because exact inversion of the Jacobian is no longer beneficial, owing
to the mismatch between Jacobian and residual. However, rigorous criteria for
determining the optimal level of convergence of the linear system at each time
step have yet to be determined. Simple iterative schemes such as Jacobi and
Gauss-Seidel schemes have been utilized successfully in the context of implicit
schemes with first-order linearizations (Struijs et al 1989, Fezoui & Stoufflet
1989, Batina 1993, Anderson et al 1995). However, since these are still lo-
cal techniques, their convergence degrades with grid size. More sophisticated
techniques such as preconditioned GMRES methods provide enhanced con-
vergence rates, particularly when applied with powerful preconditioners such
as Incomplete Lower Upper (ILU) factorization methods, which provide more
global information for the solution of the linear system (Venkatakrishnan &
Mavriplis 1993, Barth & Linton 1995, Anderson et al 1995).
Although first-order linearization methods require substantially less memory
overheads than exact linearization methods, their memory requirements are still
nontrivial. For example, on a three-dimensional tetrahedral grid, the first-order
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

UNSTRUCTURED GRID TECHNIQUES 499

Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org
by CAPES on 03/24/09. For personal use only.

Figure 9 Convergence rates of the ILU-preconditioned Newton-Krylov method, versus Gauss-

Siedel implicit method and multigrid method using implicit Gauss-Seidel as a multilevel smoother
in terms of iteration count, and CPU-time, for two-dimensional inviscid subsonic flow over NACA
0012 airfoil. Solid line: Gauss Siedel; dotted line: ILU preconditioned Newton-Krylov; dashed
line: multigrid method (using Gauss Siedel as smoother). (Reproduced from Nielsen et al 1995
with permission.)

Jacobian of a vertex-based scheme requires on the order of 350 storage loca-

tions per vertex, over three times the number required for the implementation
of an explicit scheme. An ILU preconditioning strategy can require an equiv-
alent additional amount of memory. Since GMRES methods only require the
evaluation of matrix vector products of the form ∂R
∂w
· 1w, it is possible to forgo
the storage of the Jacobian, and evaluate the Jacobian vector product directly
by finite-difference techniques:
∂R R(w + 1w) − R(w)
· 1w = (5)
∂w 
where  represents a small parameter. This requires multiple residual evalu-
ations (one for each matrix vector product) and represents a classic tradeoff
between storage and computation, particularly for expensive nonlinear residual
constructions. However, it also permits the use of the more accurate second-
order linearization in the implicit scheme, by using the same second-order
residual in the finite difference evaluation as in the right-hand side of equation
(4), as is done in the so-called Newton-Krylov methods (Cai et al 1994, Barth
& Linton 1995, Nielsen et al 1995). Figure 9 compares the convergence rates
of a Newton-Krylov method that employs ILU preconditioning with conver-
gence of a Gauss-Seidel iterative method applied to a first-order linearization,
as well as with the convergence obtained by a multigrid scheme using the same
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

500 MAVRIPLIS

Gauss-Seidel solver as a multilevel smoother, as described in section 4.3. The

depicted case involves the solution of two-dimensional subsonic inviscid flow
over a NACA 0012 airfoil on a mesh of 8578 points, using a vertex-based
second-order upwind scheme (Nielsen et al 1995). The quadratic convergence
property of the Newton-Krylov approach is evident, which achieves a very rapid
asymptotic convergence rate in terms of number of iterations. When compared
in terms of overall CPU-time, the Newton-Krylov scheme, although still supe-
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

rior to the Gauss-Seidel implicit method, is overtaken by the multigrid scheme,

due to the expense of each Newton-Krylov iteration.
Additional improvements to Newton-Krylov methods can be achieved
through continuation techniques, which provide better initial guesses (for ex-
ample by coarse to fine mesh sequencing), in order to reduce the initial phase
of poor convergence of these schemes. However, one of the drawbacks of
by CAPES on 03/24/09. For personal use only.

these methods is that they still require powerful preconditioners in order to

be effective, and the best known preconditioners are generally matrix based
(such as ILU), which entail storage requirements similar to those of the first-
order linearization. A compromise, with obvious limitations, is to use a weaker
but low-storage preconditioner, such as block-diagonal with a Newton-Krylov
method (Shakib et al 1989). For upwind discretizations based on reconstruc-
tion techniques, higher-order Jacobian vector products can be formed by mul-
tiplying the first-order Jacobian with a higher-order reconstructed vector, as
pointed out by Barth (Barth & Linton 1995). Although the method requires
the storage of the first-order Jacobian, it is exact and, unlike finite-difference
techniques, can also be applied directly to the solution of the adjoint equa-
tions, which is required in particular formulations of the design optimization
problem.
Additional approximations to the exact Jacobian can be employed to further
reduce the memory requirements of implicit schemes. For example, one may
choose to group subregions of the mesh together and only retain the terms
of the Jacobian that pertain to coupling within each region (Chan & Mathew
1994). These regions may be determined by a variety of methods and may
exhibit varying degrees of overlap. These approximations may be applied to
the Jacobian itself but are most often applied to the matrix-based preconditioners
operating on the true Jacobian (Venkatakrishnan 1994, Barth & Linton 1995,
Chan & Mathew 1994). Alternatively, the regions may be reduced to lines
through the mesh joining neighboring vertices, as in the method of linelets
(Hassan et al 1990, Martin & Löhner 1992), which attempts to approximate the
Alternate Direction Implicit (ADI) scheme of structured grids. On sequential
machines, the regions or lines can be processed individually, thus reducing the
maximum memory requirements to that of the largest region or line (although
on parallel machines no savings are realized if each region is assigned to an
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

UNSTRUCTURED GRID TECHNIQUES 501

individual processor). The numerical coupling between regions or lines is lost

in these approaches, a loss that in turn degrades overall convergence, as the
number of regions increases. However, this degradation can be minimized
by judicious choices of the regions based on the character of the problem, a
technique that has not yet been fully exploited.
If these regions are reduced to individual grid points (i.e. all implicit cou-
pling between grid points is discarded), only the diagonal block matrices of
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

the original Jacobian are retained, and the point-implicit method is obtained
(Hassan et al 1990, Thareja et al 1988). These terms represent the coupling be-
tween the various fluid dynamic equations at a grid point. While point-implicit
methods can offer increases in efficiency over regular explicit schemes, they
are also plagued by slow convergence for fine grids. Point-implicit meth-
ods are sometimes viewed as preconditioning techniques for explicit schemes,
by CAPES on 03/24/09. For personal use only.

where the point-implicit matrix is the preconditioner (Riemslagh & Dick 1993,
Morano & Dervieux 1993, Ollivier-Gooch 1995, Pierce & Giles 1996).

4.3 Multigrid Methods

Multigrid methods are effective techniques that can deliver fast convergence
rates with minimal memory overheads. The basic idea of a multigrid method
is to accelerate the solution of a set of fine grid equations by time stepping on
a sequence of fine and coarse grids using a simple explicit scheme. Multigrid
methods capitalize on the effectiveness of explicit methods in reducing high-
frequency error modes. A high-frequency error can be thought of as a solution
component that oscillates between neighboring grid points. While explicit
methods are effective at eliminating such errors, they are relatively ineffective
at reducing low-frequency or global error components. This is to be expected,
because explicit schemes rely exclusively on local information supplied through
the residual stencil.
The multigrid strategy consists in explicitly time stepping the fine grid equa-
tions until the high-frequency errors have been annihilated. The solution is then
transferred to a coarser grid, where the remaining low-frequency errors now
manifest themselves as high-frequency errors and can be effectively damped
by the same explicit method. The process is repeated recursively through a
complete sequence of grids, where each level is responsible for a particular
bandwidth of errors. When the coarsest grid of the sequence is reached, the
corrections are interpolated back to each successively finer grid. This entire
multigrid cycle is then repeated until overall convergence is achieved.
In the context of unstructured meshes, the main issue involves the generation
of the coarse levels for a multigrid scheme. Given a fine unstructured mesh,
there is no simple procedure for constructing uniformly coarser levels, as in the
structured-grid case.
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

502 MAVRIPLIS

If the fine mesh is the product of an adaptive meshing algorithm using element
subdivision, then the nested parent meshes are natural candidates for coarse
multigrid levels (Parthasarathy & Kallinderis 1994, Braaten & Connell 1996).
Although this approach can be employed effectively when multiple adaptive
levels are present, the nested multigrid strategy places conflicting demands on
the multilevel construction. On the one hand, a very coarse mesh is desired
for rapid convergence, while a high-quality fine mesh is required for solution
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

accuracy. However, repeated subdivision of a very coarse initial mesh often

results in poor quality fine meshes, thus limiting the number of levels that may
be employed. The lack of flexibility in handling problems on grids of arbitrary
origin represents an additional shortcoming of this approach.
An alternative approach to unstructured multigrid methods is to generate a
sequence of completely independent coarse and fine meshes and use linear in-
by CAPES on 03/24/09. For personal use only.

terpolation to transfer variables back and forth between the various meshes of
the sequence, within a multigrid cycle (Mavriplis 1992, 1995c; Peráire et al
1992; Leclercq 1990; Morano & Dervieux 1993; Riemslagh & Dick 1993).
The meshes may be generated using any feasible grid generation technique and
will generally be nonnested, and may even not contain any common points. The
only requirement is that they conform to the same domain boundaries. This
technique is more flexible than the nested subdivision approach, since the fine
and coarse meshes are not constrained, and may be optimized independently
for accuracy and speed of convergence respectively. Furthermore, this ap-
proach can be applied to a problem with a prespecified fine mesh. The intergrid
transfer operators can be determined in a preprocessing operation but require
smart search techniques to determine the patterns between coarse and fine grid
elements, which in principle may overlap randomly. Once these have been
determined and stored, grid transfers are simply implemented as a weighted
gather and scatter of data between coarse and fine grid arrays. Application of
this technique to a three-dimensional inviscid flow is depicted in Figures 10
and 11. The fine grid contains 804,000 points, or 4.5 million elements, and the
overset multigrid solver achieves a six-order reduction in the residual in approx-
imately 100 multigrid cycles, which represents close to an order of magnitude
savings over the single-grid scheme.
One of the drawbacks of this approach is that it still requires the user to
generate multiple grids for a single solution, which is at best tedious. With this
in mind, several efforts have been pursued to automate the process. Methods
that remove selected mesh points, using graph-based algorithms, and retriangu-
late the remaining subset of vertices have been demonstrated as effective tools
for generating coarse-level multigrid meshes (Guillard 1993, Chan & Smith
1993, Matheson & Tarjan 1994). However, for very complex geometries, the
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

UNSTRUCTURED GRID TECHNIQUES 503

Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org
by CAPES on 03/24/09. For personal use only.

Figure 10 Two grids from the sequence of grids employed in the overset mesh multigrid scheme
for the computation of transonic flow over an aircraft configuration.
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

504 MAVRIPLIS
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org
by CAPES on 03/24/09. For personal use only.

Figure 11 Computed solution displayed as Mach contours and observed convergence rates for
the single-grid, overset-multigrid, and agglomeration-multigrid algorithms for the computation of
transonic flow over an aircraft configuration.
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

UNSTRUCTURED GRID TECHNIQUES 505

generation of coarse meshes that conform to the original geometry either man-
ually or automatically, becomes a difficult task.
An alternative is provided by agglomeration multigrid methods. Based on the
control-volume formulation, these methods form coarser-level meshes by fus-
ing or agglomerating fine-level control volumes together with their neighbors
(Lallemand et al 1992, Smith 1990, Venkatakrishnan & Mavriplis 1995a). The
resulting coarse mesh contains a smaller number of larger and more complex
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

control volumes. The precise manner in which control volumes are agglom-
erated can be controlled through a graph-based algorithm similar to the vertex
removal procedure described above. The flow solver must be modified to run
on arbitrarily shaped control volumes on the coarse levels. For inviscid-flow
control-volume formulations, this presents little difficulty, since the equations
are generally discretized as fluxes over individual control-volume faces, which
by CAPES on 03/24/09. For personal use only.

can be used to build contour integrals about arbitrarily shaped control volumes.
Figure 12 depicts two coarse levels generated by the agglomeration algorithm
for the same case discussed above in the context of the overset-mesh multi-
grid approach. The observed convergence rate of the agglomeration multigrid
method is almost identical to that obtained with the overset-mesh method, as
depicted in Figure 11.
For viscous flows, the discretization of diffusion terms on arbitrarily shaped
control volumes is no longer straightforward. An algebraic interpretation of
agglomeration multigrid provides a mechanism for dealing with equations sets
that contain diffusion terms. Borrowing from the algebraic multigrid literature
(Ruge & Stüben 1987), a coarse-grid operator may be constructed by projecting
the fine-grid operator onto the space spanned by the coarse-grid basis functions.
If IhH represents the restriction or fine- to coarse-grid transfer operator, I Hh the
prolongation or coarse- to fine-grid operator, and A the fine-grid discretization
operator, then the so-called Galerkin coarse-grid operator is given by the se-
quence of operators IhH AI Hh . If A represents a finite-volume approximation
for inviscid flow on the fine grid, and restriction and prolongation are taken as
piecewise constant operators, then the Galerkin construction recovers the tradi-
tional coarse-grid finite-volume discretization discussed above. However, this
strategy can be extended in a straightforward manner for other types of govern-
ing equations, such as the Navier-Stokes equations and field-equation turbulent
models. This approach may be interpreted algebraically as summing the con-
stituent fine-grid equations within an agglomerated coarse-grid cell to form a
smaller number of coarse-grid equations. This algebraic interpretation allows
one to decouple the geometry from the solution process, thus discarding any
notion of coarse-level spatial discretizations, which enhances the robustness of
the solution process for complicated geometries. A solution of viscous turbu-
lent flow over a three-dimensional wing configuration using the agglomeration
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

506 MAVRIPLIS
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org
by CAPES on 03/24/09. For personal use only.

Figure 12 Two coarse agglomerated levels from the sequence of levels employed by the
agglomeration-multigrid method for the computation of transonic flow over an aircraft config-
uration.
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

UNSTRUCTURED GRID TECHNIQUES 507

multigrid approach is depicted in Figures 13 and 14. The convergence rate is

shown to be relatively independent of grid size by comparing the convergence
history on a coarse grid of 300,000 points with the the convergence obtained
on a finer grid of 2.4 million points (Mavriplis & Venkatakrishnan 1995a).
In general, the multigrid convergence rates achieved for viscous flow cases
are substantially slower than those achieved for inviscid cases. This can be
surmised to some degree by comparing the convergence plots of Figures 11
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

and 14. The slower convergence in the viscous flow cases is principally due to
the anisotropic stretching of the mesh in the boundary-layer and wake regions.
Unstructured multigrid methods offer the possibility of designing schemes that
are insensitive to anisotropic effects. The graph-based agglomeration or vertex-
coarsening algorithms described above can be modified to merge only those
neighboring control volumes or delete only those neighboring vertices that are
by CAPES on 03/24/09. For personal use only.

the most strongly coupled to the current point. By coarsening or agglomerating

only in the direction of strongest coupling, as determined by the magnitude of
the coefficients in the fine-grid discrete equations, rather than eliminating all
possible neighbors, coarse levels that are optimal for the problem at hand can
be generated. In the case of a stretched-boundary-layer grid, this reduces to
coarsening across the boundary layer, or to semicoarsening, which has been
used effectively in the structured-grid context to remove anisotropy-induced
stiffness (Mulder 1989, Pierce & Giles 1996). Although these techniques are
in their infancy (Morano et al 1995, Raw 1996), they offer the possibility
of automatically accounting for grid-induced as well as algebraically induced
stiffness, since they operate at the equation level.
An alternate strategy for improving the convergence of multigrid methods
is to employ a stronger smoother as the base grid solver. Any of the implicit
methods described above may be substituted for an explicit method. Of course,
most implicit methods incur additional overheads, but the benefits can justify
these additional requirements. Point-implicit methods (Morano & Dervieux
1993, Riemslagh & Dick 1993, Ollivier-Gooch 1995), Gauss-Seidel methods
(Anderson 1994), and subregion methods (Raw 1996, Chan & Mathew 1994)
have all been demonstrated as effective multigrid solvers. There is an obvious
duality between the construction of subregions for implicit methods and coars-
ening strategies for multigrid methods that should be exploited in the future to
better couple locally implicit strategies with multigrid methods.
The above discussion on multigrid methods centers around the Full Approx-
imation Storage (FAS) technique, where multigrid is applied directly to the
nonlinear form of the governing equations. Multigrid may also be employed
as a solver for the linear system arising at each time step of an implicit scheme,
or even as a preconditioner for an iterative solution strategy such as GMRES
applied to the linearized implicit system (Chan & Smith 1993, Raw 1996).
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

508 MAVRIPLIS
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org
by CAPES on 03/24/09. For personal use only.

Figure 13 Initial 300,000-point grid and coarse agglomerated level employed in the agg-
lomeration-multigrid scheme for the computation of viscous turbulent flow over a wing config-
uration.
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

UNSTRUCTURED GRID TECHNIQUES 509

Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org
by CAPES on 03/24/09. For personal use only.

Figure 14 Computed solution displayed as Mach contours on the wall and pressure contours
on the wing surface for fine 2.4 million–point (13.6 million–tetrahedra) grid and observed con-
vergence rates of agglomeration-multigrid algorithm on 300,000-point grid and 2.4 million–point
grid, demonstrating grid-independent convergence rates.
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

510 MAVRIPLIS

While these strategies may yield desirable robustness characteristics, they are
plagued by the high memory requirements of the implicit linearization, which
are avoided in the FAS approach.

5. CONCLUSIONS
An overview of the state-of-the-art in the areas of mesh generation, adaptivity,
discretizations, and solution strategies for unstructured meshes has been given.
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

An attempt has been made to outline the merits and disadvantages of the various
approaches in these areas. Other areas such as solution techniques for unsteady
flows and parallel implementations have not been addressed owing to space
limitations. Excellent survey papers in these areas, as well as more in-depth
reviews of some of the items covered in this paper, can be found in the literature
(Thompson & Weatherill 1993; Baker 1996; Venkatakrishnan 1995, 1996;
by CAPES on 03/24/09. For personal use only.

Special course 1992; Mavriplis 1995a,b).

Although unstructured mesh techniques have undergone rapid development
in the last several years, much improvement is still needed. Three-dimensional
inviscid analyses about complex geometries using unstructured meshes can be
performed routinely with present technology. However, viscous flow analyses
using the Reynolds-averaged Navier-Stokes equations are severely hindered
for large cases by seemingly insatiable CPU-time and memory requirements,
even for steady-state calculations. In particular, most cases are limited by
hardware memory costs, which have not decreased as rapidly as the cost of CPU-
cycles over the last five years. Further advances must be sought in the areas
of higher-order schemes, efficient and robust solution algorithms, improved
mesh generation strategies (particularly for highly stretched meshes and meshes
of mixed-element types), and better adaptive criteria to enable routine use of
adaptive meshing and the generation of solutions with guaranteed error bounds.

ACKNOWLEDGMENTS
Dr. V. Venkatakrishnan is acknowledged for his helpful comments in preparing
this paper. The author is also grateful to Dr. K. W. Anderson, Dr. S. Pirzadeh,
and E. Nielsen for providing figures for this paper.

Visit the Annual Reviews home page at

http://www.annurev.org.

Literature Cited

Aftosmis M, Gaitonde D, Tavares TS. 1994. structured meshes. AIAA Pap. 94-0415
On the accuracy, stability and monotonicity Anderson WK. 1994. A grid generation and
of various reconstruction algorithms for un- flow solution method for the Euler equa-
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

UNSTRUCTURED GRID TECHNIQUES 511

tions on unstructured grids. J. Comput. Phys. Cai XC, Gropp WD, Keyes DE, Tidriri
110(1):23–38 MD. 1994. Newton-Krylov-Schwarz meth-
Anderson WK, Rausch R, Bonhaus D. 1995. ods in CFD. Proceedings of the International
Implicit multigrid algorithms for incom- Workshop on Numerical Methods for the
pressible turbulent flows on unstructured Navier-Stokes Equations. Notes on Numer-
grids. Proc. AIAA CFD Conf., 12th, San ical Fluid Mechanics, Vol. 47:17–30. Braun-
Diego. AIAA Pap. 95-1740-CP schweig/Wiesbaden: Vieweg
Babushka I, Aziz AK. 1976. On the angle con- Castro-Diaz MJ, Hecht F, Mohammadi B. 1995.
dition in the finite-element method. SIAM J. Anisotropic unstructured mesh adaptation for
Numer. Anal. 13(6) flow simulations. Int. Mesh Roundtable, 4th,
Baker TJ. 1987. Three dimensional mesh gener- Albuquerque, NM, pp. 73–85
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

ation by triangulation of arbitrary point sets. Chan TF, Mathew TP. 1994. Domain decompo-
AIAA Pap. 87-1124 sition algorithms. Acta Numer. pp. 61–143
Baker TJ. 1991. Unstructured meshes and sur- Chan TF, Smith B. 1993. Domain decompo-
face fidelity for complex shapes. Proc. AIAA sition and multigrid algorithms for ellip-
CFD Conf., 10th, Honolulu, pp. 714–25. tic problems on unstructured meshes. UCLA
AIAA Pap. 91-1591-CP CAM Rep. 93-42, Dep. Math., Univ. Calif.,
Baker TJ. 1996. Mesh adaptation strategies for Los Angeles
problems in fluid dynamics. Finite Elem. Chew LP. 1989. Constrained Delaunay triangu-
Anal. Des. In press lations. Algorithmica 4:97–108
by CAPES on 03/24/09. For personal use only.

Baldwin BS, Barth TJ. 1991. A one-equation Connell SD, Braaten ME. 1995. Semi-
turbulence transport model for high Reynolds structured mesh generation for three-
number wall-bounded flows. AIAA Pap. 91- dimensional Navier-Stokes calculations.
0610 AIAA J. 33(6):1017–24
Barth TJ. 1991a. Numerical aspects of comput- Deconinck H, Hirsch Ch, Peuteman J. 1986.
ing viscous high Reynolds number flows on Characteristic decomposition methods for
unstructured meshes. AIAA Pap. 91-0721 the multidimensional Euler equations. Lect.
Barth TJ. 1991b. A three-dimensional upwind Notes Phys. 264:216–22
Euler solver for unstructured meshes. AIAA Desideri JA, Dervieux A. 1988. Compressible
Pap. 91-1548-CP flow solvers using unstructured grids. VKI
Barth TJ. 1992. Aspects of unstructured grids Lect. Ser. 1988-05:1–115
and finite-element volume solvers for the Eu- Edelsbrunner H, Tan TS, Waupotitsch R. 1992.
ler and Navier-Stokes equations. von Karman An 0(N 2 log N ) time algorithm for the min-
Inst. Lect. Ser., AGARD Publ. R-787 max angle triangulation. SIAM J. Sci. Stat.
Barth TJ, Jespersen DC. 1989. The design and Comput. 13:994–1008
application of upwind schemes on unstruc- Fezoui L, Stoufflet B. 1989. A class of im-
tured meshes. AIAA Pap. 89-0366 plicit upwind schemes for Euler simulations
Barth TJ, Linton SW. 1995. An unstructured with unstructured meshes. J. Comput. Phys.
mesh Newton solver for compressible fluid 84:174–206
flow and its parallel implementation. AIAA Frink NT. 1992. Upwind scheme for solving the
Pap. 95-0221 Euler equations on unstructured tetrahedral
Batina JT. 1993. Implicit upwind solution al- meshes. AIAA J. 30(1):70–77
gorithms for three-dimensional unstructured Frink NT. 1994. Recent progress toward a three-
meshes. AIAA J. 31(5):801–5 dimensional unstructured Navier-Stokes flow
Baum JD, Luo H, Löhner R. 1994. A new ALE solver. AIAA Pap. 94-0061
adaptive unstructured methodology for the George PL, Hecht F, Saltel E. 1991. Auto-
simulation of moving bodies. AIAA Pap. 94- matic mesh generator with specified bound-
0414 ary. Comput. Methods Appl. Mech. Eng.
Blacker TD, Stephenson MB. 1991. Paving: 33:975–95
a new approach to automated quadrilateral Guillard H. 1993. Node nested multigrid with
mesh generation. Int. J. Numer. Methods Eng. Delaunay coarsening. INRIA Rep. No. 1898
32:811–47 Gumbert C, Lohner R, Parikh P, Pirzadeh S.
Boris JP, Book DL. 1973. Flux corrected trans- 1989. A package for unstructured grid gen-
port, 1 SHASTA, a fluid transport algorithm eration and finite element flow solvers. AIAA
that works. J. Comput. Phys. 11:38–69 Pap. 89-2175
Bowyer A. 1981. Computing Dirichlet tessala- Harten A. 1984. On a class of high-reso-
tions. Comput. J. 24(2):162–66 lution total-variation-stable finite difference
Braaten ME, Connell SD. 1996. Three di- schemes. SIAM J. Numer. Anal. 21(1):1–
mensional unstructured adaptive multigrid 23
scheme for the Navier-Stokes equations. Hassan O, Morgan K, Peraire J. 1989. An adap-
AIAA J. 34(2):281–90 tive implicit/explicit finite element scheme
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

512 MAVRIPLIS

for compressible high speed flows. AIAA Pap. JD Rice, pp. 161–95. New York: Academic
89-0363 Leclercq MP. 1990. Resolution des equations
Hassan O, Morgan K, Peraire J. 1990. An im- d’Euler par des methods multigrilles con-
plicit finite element method for high speed ditions aux limites en regime hypersonique.
flows. AIAA Pap. 90-0402 PhD thesis. Dep. Appl. Math, Univ. de Saint-
Holmes DG, Snyder DD. 1988. The generation Etienne
of unstructured meshes using Delaunay trian- Lo SH. 1985. A new mesh generation scheme
gulation. Numer. Grid Gener. Comput. Fluid for arbitrary planar domains. Int. J. Numer.
Mech. Proc. Int. Conf. Numer. Grid Gener. Methods Eng. 21:1403–26
Comput. Fluid Dyn., 2nd, Miami, ed. S Sen- Löhner R. 1993. Matching semi-structured and
gupta, J Hauser, PR Eisman, JF Thompson unstructured grids for Navier-Stokes calcula-
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

Swansea, UK: Pineridge tions. AIAA Pap. 93-3348

Hughes TJR. 1987. Recent progress in the de- Löhner R, Baum JD. 1992. Adaptive h-
velopment and understanding of SUPG meth- refinement on 3D unstructured grids for tran-
ods with special reference to the compress- sient problems. Int. J. Numer. Methods Fluids
ible Euler and Navier-Stokes equations. Int. 14:1407–19
J. Numer. Methods Fluids 7:1261–75 Löhner R, Morgan K, Peraire J, Vahdati M.
Hughes TJR, Brooks A. 1979. A multidimen- 1987. Finite element flux-corrected transport
sional upwind scheme with no crosswind (FCT) for the Euler and Navier-Stokes equa-
diffusion. In Finite Element Methods for tions. Comput. Methods Appl. Mech. Eng.
by CAPES on 03/24/09. For personal use only.

Convection Dominated Flows. New York: 7:1093–1109

ASME Luo H, Baum JD, Löhner R, Cabello J. 1993.
Hughes TJR, Franca LP, Hulbert GM. 1989. A Adaptive edge-based finite element schemes
new finite element formulation for compu- for the Euler and Navier-Stokes equations on
tational fluid dynamics. VIII. Galerkin/least- unstructured grids. AIAA Pap. 93-0336
squares method for advective-diffusive sys- Marcum DL. 1995. Generation of unstructured
tems. Comput. Methods Appl. Mech. Eng. grids for viscous flow applications. AIAA
73:173–89 Pap. 95-0212
Hughes TJR, Mallet M. 1986. A new fi- Marcum DL, Weatherill NP. 1994. Unstructured
nite element formulation for CFD. IV. A grid generation using iterative point insertion
discontinuity-capturing operator for multi- and local reconnection. AIAA Pap. 94-1926
dimensional advective-diffusive systems. Martin D, Loḧner R. 1992. An implicit linelet-
Comput. Methods Appl. Mech. Eng. 58:329– based solver for incompressible flows. AIAA
36 Pap. 92-0668
Jameson A. 1994. Analysis and design of nu- Matheson L, Tarjan R. 1994. Unstructured
merical schemes for gas dynamics. I. Arti- multigrid strategies on massively parallel
ficial diffusion, upwind biasing, limiters and computers: a case for integrated design.
their effect on accuracy and multigrid conver- Hawaii. Int. Conf. Syst. Sci., 27th, 2:169–78.
gence. Int. J. Comput. Fluid Dyn. 4:171–218 New York: IEEE
Jameson A, Baker TJ, Weatherill NP. 1986. Mavriplis DJ. 1987. Solution of the two-
Calculation of inviscid transonic flow over dimensional Euler equations on unstructured
a complete aircraft. AIAA Pap. 86-0103 triangular meshes. PhD thesis. Dep. Mech.
Jameson A, Schmidt W, Turkel E. 1981. Nu- Aerospace Eng., Princeton Univ.
merical solution of the Euler equations by fi- Mavriplis DJ. 1990a. Accurate multigrid solu-
nite volume methods using Runge-Kutta time tion of the Euler equations on unstructured
stepping schemes. AIAA Pap. 81-1259 and adaptive meshes. AIAA J. 28(2):213–
Joe B. 1989. Three-dimensional triangulations 21
from local transformations. SIAM J. Sci. Stat. Mavriplis DJ. 1990b. Adaptive mesh generation
Comput. 10:718–41 for viscous flows using Delaunay triangula-
Kennon SR, Meyering JM, Berry CW, Oden tion. J. Comput. Phys. 90(2):271–91
JT. 1992. Geometry based Delaunay tetrahe- Mavriplis DJ. 1991a. Algebraic turbulence
dralization and mesh movement strategies for modeling for unstructured and adaptive
multi-body CFD. AIAA Pap. 92-4575 meshes. AIAA J. 29(12):2086–93
Lallemand M, Steve H, Dervieux A. 1992. Mavriplis DJ. 1991b. Unstructured and adaptive
Unstructured multigridding by volume ag- mesh generation for high-Reynolds number
glomeration: current status. Comput. Fluids viscous flows. Proceedings of the Interna-
21(3):397–433 tional Conference on Numerical Grid Gen-
Lawson CL. 1972. Transforming triangulations. eration: Computational Fluid Dynamics and
Discrete Math. 3:365–72 Related Fields, 3rd, Barcelona, Spain, ed. AS
Lawson CL. 1977. Software for C 1 surface in- Arcilla, J Hauser, PR Eisman, JF Thompson,
terpolation. In Mathematical Software III, ed. pp. 79–92. New York: North-Holland
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

UNSTRUCTURED GRID TECHNIQUES 513

Mavriplis DJ. 1992. Three-dimensional multi- laminar flows. Proc. 6th Int. Symp. CFD,
grid for the Euler equations. AIAA J. Lake Tahoe, NV
30(7):1753–61 Osher S. 1984. Riemann solvers, the en-
Mavriplis DJ. 1995a. Multigrid techniques for tropy condition and difference approxima-
unstructured meshes. VKI Lect. Ser. Comput. tions. SIAM J. Numer. Anal. 21(2):217–
Fluid Dyn., 26th, VKI-LS 1995-02 35
Mavriplis DJ. 1995b. A three-dimensional Paillere H, Deconinck H, Struijs R, Roe PL,
multigrid Reynolds-averaged Navier-Stokes Mesaros LM, Muller JD. 1993. Computa-
solver for unstructured meshes. AIAA J. tions of compressible flows using fluctuation-
33(3):445–53 splitting on triangular meshes. AIAA Pap. 93-
Mavriplis DJ. 1995c. Unstructured mesh gener- 3301-CP
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

ation and adaptivity. VKI Lect. Ser. Comput. Palmerio B. 1994. An attraction-repulsion mesh
Fluid Dyn., 26th, VKI-LS 1995-02 adaption model for flow solution on un-
Mavriplis DJ, Venkatakrishnan V. 1994. Ag- structured grids. Comput. Fluids 23(3):487-
glomeration multigrid for two-dimensional 506
viscous flows. Comput. Fluids 24(5):553-70 Parpia I, Michalak DJ. 1993. Grid-independent
Mavriplis DJ, Venkatakrishnan V. 1995a. A upwind scheme for multidimensional flows.
3D agglomeration multigrid solver for the AIAA J. 31(4):646–51
Reynolds-averaged Navier-Stokes equations Parthasarathy V, Kallinderis Y. 1994. New
on unstructured meshes. AIAA Pap. 95-0345 multigrid approach for three-dimensional
by CAPES on 03/24/09. For personal use only.

Mavriplis DJ, Venkatakrishnan V. 1995b. A uni- unstructured, adaptive grids. AIAA J.

fied multigrid solver for the Navier-Stokes 32(5):956–63
equations on mixed element meshes. AIAA Peraire J, Peirö J, Morgan K. 1992. A 3D finite-
Pap. 95-1666 element multigrid solver for the Euler equa-
Merriam ML. 1991. An efficient advancing- tions. AIAA Pap. 92-0449
front algorithm for Delaunay triangulation. Peraire J, Vehdati M, Morgan K, Zienkiewicz
AIAA Pap. 91-0792 OC. 1987. Adaptive remeshing for compress-
Morano E, Dervieux A. 1993. Looking for O(N) ible flow computations. J. Comput. Phys.
Navier-Stokes solutions on non-structured 72:449–66
meshes. Copper Mountain Conf. Multigrid Pierce N, Giles M. 1996. Preconditioning on
Methods, 6th, pp. 449–64. NASA CP-3224 stretched meshes. AIAA Pap. 96-0889. J.
Morano E, Mavriplis DJ, Venkatakrishnan V. Comput. Phys. Submitted
1995. Coarsening strategies for unstruc- Pirzadeh S. 1994a. Unstructured viscous grid
tured multigrid techniques with application to generation by the advancing-layers method.
anisotropic problems. ICASE Rep. No. 95-34. AIAA J. 32(8):1735–37
ICASE, NASA Langley Res. Cent., Hamp- Pirzadeh S. 1994b. Viscous unstructured three-
ton, VA. SIAM J. Sci. Stat. Comput. Sub- dimensional grids by the advancing-layers
mitted method. AIAA Pap. 94-0417
Mortenson ME. 1985. Geometric Modeling. Preparata FP, Shamos MI. 1985. Computational
New York: Wiley & Sons Geometry, An Introduction. Texts Monogr.
Mueller JD, Roe PL, Deconinck H. 1992. A Comput. Sci. New York: Springer-Verlag
frontal approach for node generation in De- Rausch RD, Batina JT, Yang HTY. 1992. Spa-
launay triangulations. VKI Lect. Ser., AGARD tial adaptation of unstructured meshes for
Publ. R-787 unsteady aerodynamic flow computations.
Mulder WA. 1989. A new multigrid approach AIAA J. 30(5):1243–51
to convection problems. J. Comput. Phys. Raw M. 1996. Robustness of coupled algebraic
83(2):303–23 multigrid for the Navier-Stokes equations.
Nakahashi K. 1987. FDM-FEM approach for AIAA Pap. 96-0297
viscous flow computations over multiple bod- Rebay S. 1993. Efficient unstructured mesh gen-
ies. AIAA Pap. 87-0604 eration by means of Delaunay triangulation
Ni RH. 1982. A multiple-grid scheme for and Bowyer/Watson algorithm. J. Comput.
solving the Euler equations. AIAA J. Phys. 106(1):125–38
20(11):1565–71 Riemslagh K, Dick E. 1993. A multigrid method
Nielsen EJ, Anderson WK, Walters RW, Keyes for steady Euler equations on unstructured
DE. 1995. Application of Newton-Krylov adaptive grids. See Morano & Dervieux
methodology to a three-dimensional unstruc- 1993, pp. 527–42
tured Euler code. Proc. AIAA CFD Conf., Roe PL. 1981. Approximate Riemann solvers,
12th, San Diego. AIAA Pap. 95-1733-CP parameter vectors and difference schemes. J.
Ollivier-Gooch CF. 1995. Towards problem- Comput. Phys. 43(2):357–72
independent multigrid convergence rates for Roe PL. 1986. Discrete models for the nu-
unstructured mesh methods. I. inviscid and merical analysis of time-dependent multidi-
December 2, 1996 16:42 Annual Reviews Chapter14 AR23-14

514 MAVRIPLIS

mensional gas dynamics. J. Comput. Phys. upon a Voronoi type method. See Mavriplis
63:458–76 1991b, pp. 93–103
Rostand P, Stoufflet B. 1988. TVD schemes to van Leer B, Tai CH, Powell KG. 1989. Design
compute compressible viscous flows on un- of optimally-smoothing multi-stage schemes
structured meshes. Notes on Numerical Fluid for the Euler equations. AIAA Pap. 89-1933
Mechanics, Vol. 24:510–520. Braunschweig, Venkatakrishnan V. 1994. Parallel implicit
Ger: Vieweg unstructured grid Euler solvers. AIAA J.
Ruge JW, Stüben K. 1987. Algebraic multigrid. 32(10):1985–91
Multigrid Methods, SIAM Frontiers in Ap- Venkatakrishnan V. 1995a. Implicit schemes
plied Mathematics, ed. SF McCormick, pp. and parallel computing in unstructured grid
73–131. Philadelphia: SIAM CFD. VKI Lect. Ser. VKI-LS 1995-02
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

Shakib F, Hughes TJR, Johan Z. 1989. A multi- Venkatakrishnan V. 1995b. On the accuracy of
element group preconditioned GMRES al- limiters and convergence to steady state so-
gorithm for nonsymmetric problems arising lutions. J. Comput. Phys. 118:120–30
in finite element analysis. Comput. Methods Venkatakrishnan V. 1996. A perspective on
Appl. Mech. Eng. 87:415–56 unstructured grid flow solvers. AIAA J.
Sidilkover D. 1994. A genuinely multidimen- 34(3):533–47
sional upwind scheme and efficient multigrid Venkatkrishnan V, Barth TJ. 1989. Application
solver for the compressible Euler equations. of direct solvers to unstructured meshes for
ICASE Rep. No. 94-84. ICASE, NASA Lan- the Euler and Navier-Stokes equations using
by CAPES on 03/24/09. For personal use only.

gley Res. Cent., Hampton, VA upwind schemes. AIAA Pap. 89-0364

Slack DC, Whitaker DL, Walters RW. 1994. Venkatakrishnan V, Mavriplis DJ. 1993. Im-
Time integraton algorithms for the two- plicit solvers for unstructured meshes. J.
dimensional Euler equations on unstructured Comput. Phys. 105(1):83–91
meshes. AIAA J. 32(6):1158–66 Venkatakrishnan V, Mavriplis DJ. 1995a.
Smith WA. 1990. Multigrid solution of tran- Agglomeration multigrid for the three-
sonic flow on unstructured grids. Recent Ad- dimensional Euler equations. AIAA J.
vances and Applications in Computational 33(4):633–40
Fluid Dynamics, pp. 93–103. Proc. ASME Venkatakrishnan V, Mavriplis DJ. 1995b. Im-
Winter Annu. Meet., ed. O Baysal plicit method for the computation of unsteady
Spalart PR, Allmaras SR. 1992. A one-equation flows on unstructured grids. Proc. AIAA CFD
turbulence model for aerodynamic flows. Conf., 12th, San Diego. AIAA Pap. 95-1705-
AIAA Pap. 92-0439 CP
Special course on unstructured grid methods for Ward S, Kallinderis Y. 1993. Hybrid pris-
advection dominated flows. 1992. AGARD matic/tetrahedral grid generation for complex
Publ. R-787 3D geometries. AIAA Pap. 93-0669
Stoufflet B, Periaux J, Fezoui F, Dervieux A. Watson DF. 1981. Computing the n-
1987. Numerical simulation of 3-D hyper- dimensional Delaunay tessalation with
sonic Euler flows around space vehicles using application to Voronoi polytopes. Comput. J.
adapted finite-elements. AIAA Pap. 87-0560 24(2):167–71
Struijs R, Vankeirsbilck P, Deconinck D. 1989. Weatherill NP. 1985. The generation of un-
An adaptive grid polygonal finite-element structured grids using Dirichlet tessalations.
method for the compressible flow equations. Princeton Univ. Dep. Mech. Aerospace Eng.
AIAA Pap. 89-1957-CP Rep. MAE 1715
Tembulkar JM, Hanks BW. 1992. On generat- Weatherill NP, Hassan O, Marcum DL. 1993.
ing quadrilateral elements from a triangular Calculation of steady compressible flowfields
mesh. Comput. Struct. 42(4):665–67 with the finite-element method. AIAA Pap.
Thareja RR, Stewart JR, Hassan O, Morgan K, 93-0341
Peraire J. 1988. A point implicit unstructured Yerry MA, Shephard MS. 1984. Automatic
grid solver for the Euler and Navier-Stokes three dimensional mesh generation by the
equations. AIAA Pap. 88-0036 modified-octree technique. Int. J. Numer.
Thompson JF, Weatherill NP. 1993. Aspects of Methods Eng. 20:1965–90
numerical grid generation: Current science Zhu JZ, Zienkiewicz OC, Hinton E, Wu J. 1991.
and art. AIAA Pap. 93-3539-CP A new approach to the development of auto-
Vallet MG, Hecht F, Mantel B. 1991. matic quadrilateral mesh generation. Int. J.
Anisotropic control of mesh generation based Numer. Methods Eng. 32:849–66
 Annual Review of Fluid Mechanics
Volume 29, 1997

CONTENTS
G. I. TAYLOR IN HIS LATER YEARS, J. S. Turner 1

ELECTROHYDRODYNAMICS: The Taylor-Melcher Leaky Dielectric

Model, D. A. Saville 27
CORE-ANNULAR FLOWS, D. D. Joseph, R. Bai, K. P. Chen, Y. Y.
Renardy 65

CONVECTION IN MUSHY LAYERS, M. Grae Worster 91

QUANTIFICATION OF UNCERTAINTY IN COMPUTATIONAL
FLUID DYNAMICS, P. J. Roache 123
Annu. Rev. Fluid Mech. 1997.29:473-514. Downloaded from arjournals.annualreviews.org

COMPUTING AERODYNAMICALLY GENERATED NOISE, Valana

L. Wells, Rosemary A. Renaut 161
NONLINEAR BUBBLE DYNAMICS, Z. C. Feng, L. G. Leal 201

PARABOLIZED STABILITY EQUATIONS, Thorwald Herbert 245

by CAPES on 03/24/09. For personal use only.

QUANTITATIVE FLOW VISUALIZATION IN UNSEEDED FLOWS,

Richard B. Miles and, Walter R. Lempert 285
CONVECTION INTO DOMAINS WITH OPEN BOUNDARIES, T.
Maxworthy 327

FLUID MECHANICS OF SPIN CASTING OF METALS, Paul H.

Steen, Christian Karcher 373
BLOOD FLOW IN ARTERIES, David N. Ku 399

THE PHENOMENOLOGY OF SMALL-SCALE TURBULENCE, K. R.

Sreenivasan, R. A. Antonia 435

UNSTRUCTURED GRID TECHNIQUES, D. J. Mavriplis 473

MODERN HELICOPTER AERODYNAMICS, A. T. Conlisk 515